Assessment of DRAINMOD-NII Model for Prediction of Nitrogen Losses Through Subsurface Drained Sandy Clay Under Cultivation in South West Punjab, India

9 In the present study, DRAINMOD-NII model was calibrated for the years 2018-2019 and validated 10 for the period 2019-2020 over the two cropping years. The model simulations were statistically evaluated 11 by comparing the measured drain flows and nitrate-nitrogen (NO 3 -N) with the model simulated drain 12 outflows and nitrate loss. The study results depicted closer agreement between the simulated and observed 13 results for both the calibration and validation periods. The Root Mean Square Error (RMSE) of the drainage 14 rate was 8.88 cm more than observed data,15.41, 0.53 and 0.57 cm were the values recorded for PBIAS, 15 modelling efficiency (NSE) and R 2 . The similar parameter values for nitrogen load were recorded to be 16 0.14, 2.76 ,0.84 and 0.88 respectively during the calibration period for rice wheat system. The model was 17 statistically tested during the validation period also, confirming DRAINMOD-NII has the capability to 18 simulate nitrogen losses from the area subjected to subsurface drainage system


Introduction 27
A major source of surface and groundwater pollution has been attributed to drained agricultural 28 lands (Stoate et al., 2001;Randall and Mulla, 2001). Extensive use of manure and fertilizers to boost the 29 production of food production can increase the risk of nitrogen (N) contamination of surface water and 30 groundwater, promoting eutrophication (Carpenter et al., 1998).Nitrate-nitrogen (NO3-N) polluted drainage 31 waters have been described as a key non-point source of surface water contamination (Jacobs and Gilliam, simulation. Precipitation data was obtained from Bathinda weather station, Punjab Agricultural University. 137 Irrigation was applied to the crops on weekly basis for rice and after a fortnight for wheat crop. Potential 138 evapotranspiration (PET) depends on net radiation, wind velocity and humidity within the region. Daily 139 PET was computed, using Thornthwaite method (1948), in the model using weather data from 2018-140 2020.Soil samples were collected for various soil parameters and nitrate content form different locations 141 using gps based, sampling points, considering the latitude and longitude of that particular point. The 142 samples were collected from various depths to a maximum depth of 1.8 m, between the laterals. Hydraulic 143 conductivity was measured insitu, during both the rice and wheat seasons, using augur hole method and an 144 average representative value was selected for the whole study area. Soil properties at the study area are 145 listed in Table 2. DRAINMOD-NII simulates cropping systems, comprising more than one crop. The 146 simulation study was based on rice wheat cropping rotation, which takes almost one year to complete. The 147 study was repeated for the two years having rice wheat cropping system. The input data for each crop 148 consist of the major dates of planting, the effective rooting depth, harvesting and stress counting parameters. 149 "Crop parameters included N uptake and yield parameters. Yield parameters of harvest index (HI), 150 root/shoot ratio (RSR), N content of plant grains, roots, shoots and potential crop yield were included in the 151 model. The N uptake during the entire growing season is estimated from yields parameters by 152 DRAINMOD-N II model. The maximum crop yield obtained in absence of soil water related stresses is 153 defined as the potential crop yield by Evans et al. (1991). In DRAINMOD the crop yield was calculated 154 based on the product of potential yield and the DRAINMOD predicted relative yield. The ratio of crop yield 155 to the total above-ground biomass is defined as the Plant HI by Hay (1995). Hoad et al. (2001) described 156 the RSR as the mass ratio between root dry matter and shoot dry matter. The HI and RSR are used by the 157 DRAINMOD-N II model to estimate nongrain above-ground dry matter and below-ground dry matter from 158 DRAINMOD-N II predicted or field-measured crop yields (Youssef et al., 2005). Table 3 shows the 159 possible yields and nitrogen content of rice wheat based on field measurements. Salazar et al. (2009) listed 160 the popular ranges of rice wheat crop N, C, and lignin contents as in Table 3. The N-uptake tabulated feature 161 proposed by Youssef et al. (2003) and Shedekar et al. (2021) was used for crop modeling in addition to the harvest index, root/shoot ratio, shoot N, and root N in rice wheat that were calculated based on observed 163 field data." 164   Denitrification, nitrification, fertilizer dissolution, pH regulation, and volatilization were among the carbon 170 and nitrogen transformation parameters considered in the model simulation. Organic matter parameters 171 define the possible rates of decomposition (Kdec) and C/N ratios organic matter and litter in soils (SOM) 172 pools. The model was initialized with NO3-N, NH4-N, and Organic Carbon (OC) concentrations measured in the region. The model was simulated using the procedure defined by Youssef et al. (2006). The nitrogen 174 initial transport parameters and NH4 distribution coefficient input parameters for DRAINMOD-NII are 175 shown in Table 4. Table 5 shows the values chosen during model calibration. All of the values mentioned 176 were derived from Salazar et al. (2009) and Shedekar et al. (2021). 177 3.5 0.5

Transformation parameters for Soil Organic Carbon
"Active pool decomposition rate Passive pool decomposition rate (day − 1 )" 6.2E -6 to 24.7E -6 12.3288E -5 Oi represents the observed value at time i, Pi represents the simulated value at time i. Obar represents the 190 observed mean value, and n represents the number of paired observed-simulated values. When the RMSE 191 (Equation (1)) equals 0 (zero), it implies a perfect match between observed and expected values, whereas 192 increasing RMSE values imply a worsening match. According to Singh et al. (2004), RMSE values less 193 than half the standard deviation of the observed (measured) data are considered low and suggest a strong 194 model prediction. Nash-Sutcliffe efficiency (Equation (2)) can range from −∞ to 1, according to Nash and 195 Sutcliffe (1970). An efficiency of 1 (E = 1) corresponds to a perfect match between the simulated and 196 observed results. When the efficiency is zero (E = 0), the model predictions are as accurate as the mean of 197 the observed results. Efficiency less than zero (E< 0) means that the observed mean is a better predictor 198 than the model. The coefficient of determination, R 2 , (Equation (3)), ranges from 0 to 1 that defines how 199 much of the variance in the calculated data is explained by the model, with higher values meaning less error 200 variance, R 2 > 0.5 is usually considered appropriate (Santhi et al., 2001;Van Liew et al., 2003). The average 201 tendency of the simulated data to be greater or smaller than their observed counterparts is measured by the 202 percentage of bias (PBIAS) (Gupta et al., 1999). PBIAS optimal value is 0, with low magnitude values 203 suggesting a good model simulation. Underestimation bias is indicated by positive values, while 204 overestimation bias is indicated by negative values (Gupta et al., 1999). The results of a statistical analysis 205 were used to verify the model's reliability during the calibration and validation periods, and the results are 206 shown in Table 6 and 7. The significance of R 2 was determined using a partial F test, which revealed a 207 substantial correlation at a 5% level of significance, as shown in Table 6 and 7 and discussed in sections performance to be good (Moriasi et al., 2007) to excellent . The predicted and observed 220 In both years, the greatest discrepancies between the model simulated and observed monthly drain outflows 225 occurred in March, June, and July. There was less rainfall during this period, but more drainage was 226 observed. The model appears to be more sensitive to rainfall than the observed data suggests. The expected 227 monthly drain outflows were, however, in good agreement with the observed values during all other 228 periods. Overall,48% of the outflow from the study area was recorded in the months of Jan, March, July, 229 August and December. The comparison between the observed and predicted cumulative drain outflows 230   In the calibration period, the average of simulated daily drainage discharges during the rice wheat 242 season (0.88 cm day −1 ) was closer to that observed (1.02 cm day −1 ). Flow events during the growing season 243 are predicted well, but there are discrepancies in the magnitude of some events, especially during wet 244 months, such as June and July, when drainage was under-estimated. Since, subsurface drainage and 245 evapotranspiration are the two main pathways of water loss considered in this simulation study's water 246 balance, any over-or under-estimation of subsurface drainage is balanced by adjustments in simulated ET. 247 Actual ET is calculated using daily potential ET and soil moisture availability within the crop rooting depth 248 by the DRAINMOD model. Daily potential ET was calculated using the Thornthwaite equation and crop coefficients by (Allen et al., 1998) as input to the model in this modeling research. Actual crop water use 250 varies on year-to-year basis depending on moisture availability, and crops can change their water use by 251 their rooting depth. In the summers of 2018 and 2019, simulated drain flow depths matched well with the 252 observed drain flow depths, but drainage was underestimated in June and July. This implies that better ET 253 estimation will help the model perform better. However, this is only possible if more accurate weather data 254 or calculated crop ET are available at the desired location. 255

Simulated nitrate loads 256
In many studies, the movement of NO3-N has been linked to the movement of water in agricultural 257 soils (Armstrong and Burt, 1993). Figure 6 depicts the effects of daily and accumulated NO3-N losses in 258 subsurface drainage. Table 6 shows the statistical indices estimated for the predicted and observed daily 259 NO3-N losses. The calibration for the study area appears to be satisfactory, and the total expected NO3-N 260 losses are in good agreement with the observed values, as shown in Fig. 6. The cumulative predicted NO3-261 N losses of 32.15 kg ha −1 was recorded during the growing season in subsurface drainage over the 262 calibration period which was 2.9% less than the observed NO3-N losses of 33.07 kg ha −1 . King et al. (2015) 263 analyzed the effects of crop type and season on subsurface drainage discharge and nutrient loads. 264 " Bjorneberg et al. (1996) reported over a three-year period in Iowa that up to 85% of annual tile flow and 265 NO3-N loads from a corn-soybean and continuous corn rotation on loam soils occurred during the non-266 growing seasons. The averages of daily simulated and measured NO3-N losses were, 1.53 and 1.57 kg ha -267 1 , respectively during calibration (Table 8) under the rice wheat growing season. The NO3-N losses in drain 268 outflows were strongly dependent on outflow rates in the study area. The calibrated and validated daily 269 drain flows were significantly different than their corresponding simulated values (Table 8)

Simulated drainage outflows 290
The experimental plot data was used to validate the calibrated model DRAINMOD-N II for the 291 period (2019-2020). As shown in Figs.7 (a and b), the results visually indicate strong agreement between 292 the field measurements and the simulated values for monthly drainage outflows. The comparisons were 293 tested using t test, whereby the differences between observed and simulated drainage outflows were non-294 significant for the year 2018 while as significant difference was found in the years 2019 and 2020, 295 respectively. The DRAINMOD-NII model predicted values of monthly subsurface drainage outflows 296 matched well with the observed data, during the validation period, except in the months of March, June and 297 July of 2019 and 2020 respectively (Fig.7. a and b). This may be due to, differences in precipitation between 298 the rain gauge site and experimental plots, which might probably, have affected predictions in March 299 (having observed subsurface drain outflow of 36.51 cm and precipitation 1.06 cm), June and July (having 300 observed subsurface drain outflow of 35.36, 36.51 and precipitation of 3.2 and 39.74 cm, respectively). In 301 March, June of 2020 also, the observed subsurface drain outflows were (36.50 and 35.34 cm, respectively) 302 which was unrealistically very high (79% and 88%) when compared to precipitation (7.33 cm and 4.2 cm 303 respectively). Overall, the model performance during the validation period was "good" with NSE = 0.47; 304 PBIAS=19.52; RMSE=9.66; R 2 = 0.53 (Table 7). The predicted monthly subsurface drainage values were 305 more variable compared to observed data, as indicated by larger than calibrated PBIAS values in case of 306 simulated data (Table 7). The cumulative predicted subsurface drainage outflow of 296.57 cm was recorded 307 over the entire validation periods, which was 20% lower than the observed subsurface drain outflow of 308 376.45 cm. The average of simulated daily drainage discharges was 0.84 cm day −1 during the rice-wheat 309 cropping season which was closer to that of the observed 1.08 cm day −1 , during the validation period. 310 During the two growing seasons, there were more discrepancies between simulated and observed drain 311 discharges in both the calibration and validation processes. The major reason for such discrepancy in the 312 model can also be related to the differences in soil conditions and rainfall amount and pattern in different growing seasons. After a precipitation or irrigation event, the infiltrated water is used to replenish the 314 moisture lost from the dry zone first, and the remaining portion joins the excess soil moisture, eventually 315 causing a rise in water table. An under-estimation of drainage outflow may occur when a portion of 316 precipitation was used to replenish the dry zone, resulting in lower water table rise than expected. During 317 rainfall events, model predictions were usually lower than observed values, suggesting that simulated 318 effects were more unpredictable. "Due to the existence of a hard pan layer below the plough pan of the 319 study field (Darzi-Naftchali et al., 2013), part of the excess water travels horizontally in the plough layer to 320 a drain trench, then vertically to a drain pipe. The vertical hydraulic conductivity or vertical resistance of 321 the soil layer beneath the plough pan significantly influences flow to drains under these conditions, which 322 was not taken into account in DRAINMOD simulations. After evaluating uncertainties related to 323 DRAINMOD predictions for drainage outflows, Wang et al. (2006) concluded that vertical saturated 324 hydraulic conductivity of the restrictive layer is the most sensitive parameter. The different paths of seepage 325 were not included in the current analysis, with the exception of vertical seepage, which forms a major part 326 of the water balance. It is said to have a significant impact on the DRAINMOD model predictions (Chang 327 et al., 1983), which depend on exact quantification of all water balance components." 328 Table 7

Simulated nitrate loads 339
The values of statistical parameters comparing simulated and observed NO3-N losses were lower 340 in validation period than in calibration (Table 7). The PBIAS value was -5.19 kg N ha -1 , the modelling 341 efficiency (NSE) was 0.82 and R 2 0.85. The slightly lower agreement shown by statistical parameters for 342 validation when compared with calibration period was caused by a few events with larger discrepancies 343 between the simulated and observed NO3-N losses (Fig. 8). The majority of these discrepancies occurred 344 in the months of July, November and January (Table 8), in the remaining months, predicted NO3-N losses 345 in subsurface drainage matched well with the observed losses. The predicted NO3-N losses in subsurface 346 drainage were much less variable compared to observed data (Fig. 8). The cumulative predicted NO3-N 347 losses of 28.77 kg ha −1 was recorded in subsurface drainage over the validation period which was 6.5% 348 more than the observed NO3-N losses of 26.88 kg ha −1 . The model might have under predicted 349 denitrification during the validation period, leaving more mineral N susceptible to leaching in the drainage 350 outflow (Fig.8). It is possible that the overprediction of model for N mineralization rates, could partly 351 contribute to the errors in predicting NO3-N losses with drain outflow during these periods. Neither N 352 mineralization nor denitrification was possible to measure, in the field to test the accuracy of model 353 prediction for these quantities. However, the value of these parameters was taken from the literature rather 354 than field and laboratory values that may cause some errors in predicting monthly NO3-N drainage losses 355 using DRAINMOD-NII model.

Conclusions
DRAINMOD-NII was successfully calibrated and validated over a two-year cultivation period in Thehri, Muktsar, Punjab, using data sets from conventional drained plots. The statistical comparison of simulated and observed drain flows and nitrogen losses revealed a stronger agreement with some discrepancies between the two data sets. For the DRAINMOD-N II, statistical goodness-of-fit measurements such as root mean square error (RMSE), percentage of bias (PBIAS), modeling efficiency (NSE), and coefficient of determination (R 2 ) confirm that the model results and field observations are in strong agreement. The findings show that DRAINMOD-N II has the ability to simulate drainage rates and nitrogen losses from agricultural lands in Indian conditions for newly reclaimed lands. However, since this model was only tested for a shorter time period and with a single cropping method, inconsistencies that caused conflict with a specific set of values must be resolved for wider application of this model.